🤚 The Open-Palm Oracle

There was a time — roughly eighteen months ago, which in AI years qualifies as the Mesozoic Era — when the smartest money in the room was the room itself. Prediction markets like Polymarket and Kalshi were the gold standard for forecasting real-world events: elections, economic indicators, sporting upsets, whether your favorite AI company would survive its next funding round. The wisdom of the crowd, liquid, tradeable, and smugly efficient.

Then the machines showed up, placed their bets, and won.

According to research from the University of Chicago’s SIGMA Lab, AI models can now forecast real-world events as accurately as prediction markets — and in several cases, better. Their platform, Prophet Arena, launched in August 2025, pits large language models against live market odds on unresolved events. The results are, to use a technical term, uncomfortable for humans.

GPT-5 currently leads with a Brier score of 82.21%. OpenAI’s o3-mini posted the highest simulated profit returns — in one case, correctly predicting Toronto FC would beat San Diego FC when the market gave Toronto just 11% odds. The model pegged it at 30%. Toronto won. The model collected its hypothetical nine dollars. The market collected its hypothetical humility.

👐 The Two-Handed Reckoning

But wait — because there’s always a “but” when AI does something impressive, and it usually involves humans trying to feel better about themselves.

Prediction markets aren’t just pattern-matching machines. They aggregate incentivized human judgment: real money, real stakes, real consequences for being wrong. When a Polymarket trader puts $10,000 on a political outcome, they’re not running a probability calculation — they’re channeling gossip, insider context, gut instinct, and the kind of motivated reasoning that only money can buy.

AI, meanwhile, is reading the internet and connecting dots. It doesn’t have access to the whispered conversations at Davos. It doesn’t know that a senator’s chief of staff just updated their LinkedIn. What it does have is the ability to synthesize thousands of data points in seconds and arrive at a probability estimate without emotional bias, recency effects, or a hangover.

The real kicker? PredIQt, a platform launched in January 2026 by IQ AI, took this a step further — deploying autonomous AI agents to actually trade on Polymarket with real positions. In their first completed season, a Claude-based agent called Kassandra delivered a 29% return over 17 days. The Gemini-based agent managed 12%. The GPT-based agent — bless its heart — posted a 19% loss. We will not be making jokes about this. (We absolutely will.)

🌿 The Gentle Awakening

Here is the part where we step back from the spreadsheets and acknowledge what’s actually happening: the concept of “market intelligence” is being unbundled.

For decades, the efficient market hypothesis told us that prices reflect all available information. Prediction markets extended this logic to events: if enough people bet on an outcome, the aggregate odds are the best forecast available. It was elegant. It was democratic. It was, apparently, improvable.

What the University of Chicago research suggests isn’t that prediction markets are broken — it’s that they’re slow. Markets need liquidity. They need participants. They need time to converge. An AI model can read a breaking news story, cross-reference it against seventeen databases, and update its probability estimate before a single trader has finished reading the headline.

The implications for institutional decision-making are staggering. Insurance companies, hedge funds, government agencies, logistics firms — anyone who currently pays humans to predict the future should be watching this research with the kind of attention usually reserved for fire alarms.

Of course, the models also diverge wildly on certain questions. On federal AI regulation, the market sat at 25% probability. Qwen 3 said 75%. GPT-4.1 said 60%. Llama 4 said 35%. When the oracles disagree, the customer is left choosing which robot to trust — which, ironically, is just prediction markets with extra steps.

👑 The Crown Verdict

We are now living in an era where AI doesn’t just analyze markets — it participates in them. It doesn’t just summarize probability — it generates it. And in at least one documented case, it does so more profitably than a crowd of financially motivated humans.

This is not the singularity. This is something arguably more unsettling: the commodification of foresight. When a language model can outperform a prediction market, the question isn’t whether humans are obsolete — it’s whether crowds are. The wisdom of the crowd may soon be replaced by the confidence interval of the cluster.

Peter Diamandis, who highlighted this development in his MOONSHOTS series, has been saying for years that exponential technologies feel absurd in real time. He’s right. An AI agent named Kassandra — named after the prophet cursed to never be believed — just posted a 29% return. The irony writes itself, and honestly, an AI probably wrote it better.

Inspired by AI Just Beat Prediction Markets | MOONSHOTS by Peter Diamandis.

Your forecast is showing. Hedge wisely.

“The market priced it at eleven percent. The machine priced it at thirty. The machine was right, but the market still charged a fee.” — The Slap of Wisdom Quantitative Divination Desk, currently rebalancing its portfolio of existential dread